Six days. That’s how long Anthropic took to ship Opus 5.5 and Sonnet 5.5 back to back, September 22 to September 28, 2026. Haiku 5.5 was named in the same breath as both of them and has shipped nothing at all. No date, no price, no model card, no benchmark numbers. Just a confirmation that it exists and the word “soon.”
I review toolkits for a living, which means I spend most of my week deciding what to actually plug into a pipeline versus what to wait on. Haiku 5.5 is currently in the second bucket, and I want to explain why that matters more than the hype cycle suggests.
What’s actually on the table right now
The Claude lineup as of late September 2026 looks like this: Fable 5.1 at $10/$50 per million tokens, Opus 5.5 at $4/$20, Sonnet 5.5 at $2/$10, and Haiku 4.5 at $1/$5. Haiku 4.5 shipped October 15, 2025. That makes it the oldest model in a family where everything above it got refreshed inside a single week.
If you’re building anything high-volume — classification, routing, extraction, cheap first-pass summarization — Haiku is the tier you live in. It’s the one where cost per call actually shows up on the invoice. And it’s the one Anthropic has left sitting on a generation-old base while the flagship tiers leapfrogged twice.
Why the gap is the story
Opus 5.5 and Sonnet 5.5 both carry a 1M token context window. Haiku 4.5 predates that shift entirely. So anyone running a mixed-model setup right now is dealing with a real architectural mismatch: your expensive models can hold an entire codebase, and your cheap model can’t. That forces chunking logic, routing rules, and fallback paths that exist purely because one tier is behind.
I’ve seen teams solve this the expensive way — route everything to Sonnet because the context math is simpler. That works until the bill arrives. A $1/$5 model doing 80% of your volume versus a $2/$10 model doing the same work is a 2x difference on the line item that scales fastest.
My honest read on the silence
Anthropic shipped two models in under a week. That’s not a company struggling to execute. So the absence of a Haiku 5.5 date reads less like a delay and more like a deliberate sequencing choice. Small models are harder to get right in a specific way: you’re distilling capability into a budget, and the tradeoff curve is unforgiving. Ship a small model that benchmarks poorly and you’ve undercut the whole family’s credibility at the exact tier where most real traffic lives.
There’s also no model card. For reviewers and for anyone doing procurement, that’s the actual blocker. I can’t tell you whether Haiku 5.5 is worth waiting for because there is nothing to evaluate. No eval scores, no context window figure, no pricing. Anyone writing confident takes about its performance right now is making it up.
What I’d do if I were building this week
Practical guidance, based on what’s shippable today rather than what’s rumored:
- Don’t architect around a model that doesn’t exist. Build with Haiku 4.5 and keep your model identifier in config, not hardcoded. Swapping later should be a one-line change.
- Test your routing assumptions on context, not just quality. If your cheap-tier calls are bumping context limits, that’s a design problem you’ll have regardless of what ships next.
- Price your pipeline at both tiers. Run the math at $1/$5 and at $2/$10 so you know how much a forced upgrade would cost you. If the answer is painful, that’s a reason to keep your prompts tight now.
- Treat “due soon” as unscheduled. No date means no date. Plenty of confirmed models have slipped.
Where this leaves the toolkit
The Claude 5.5 family is in an awkward half-shipped state. Two tiers are current and strong, one tier is a year old, and the pricing ladder has a visible gap in it. That’s a normal phase of any staged rollout, but it’s a genuinely annoying one if your product economics depend on the bottom rung.
My take: Haiku 5.5 is the most consequential unreleased model in the family, precisely because nobody’s writing breathless threads about it. Flagship models win demos. Cheap models win deployments. When Anthropic finally puts a date and a price on this one, that’s the number I’ll be checking first — and the model card I’ll actually read line by line before recommending anything.
Until then, Haiku 4.5 is what you’ve got. Build accordingly, keep your swap path clean, and ignore anyone claiming to know how the new one performs.
🕒 Published: